三维运动目标的多尺度智能递推识别新方法

在实际成像条件下,运动中的三维目标,其投影形状(Silhouette)是变化的,因而其可识别性也处于变动中．为了应对这类困难情况,本文定义了模式的动态特征空间和模式的动态可识别性等概念．讨论了建立多尺度三维目标特性视图特征模型的必要性,以及将目标运动特性一般约束用于目标序列图像识别的合理性．据此,提出了处理三维目标运动图像序列的多尺度智能递推识别方法(MUSIRR)．构造了一种混合神经网络和逻辑决策模块的智能识别器,BP神经网和RBF网用作识别器的基本构成单元．在训练阶段,该识别器使用目标的多尺度二值特性视图模型的规则矩不变量为样本特征向量．在识别阶段,算法在递推识别序列目标图像过程中,充分利用了目标姿态不会突变以及有关成像过程的合理约束,达到了提高识别率目的．与文献中的基于单尺度特性视图的三维目标识别方法相比,本文的方法训练过程简单,只需较少的目标特性视图模型样本,不仅能处理单帧图像,更能有效处理序列图像．对几类飞机目标的大规模模拟实验结果证实了本文方法的合理性和有效性．

A Novel Multi-Scale Intelligent Recursive Recognition Method for Three-Dimensional Moving Targets

In practical imaging condition, the silhouette of a 3D moving target is changing, therefore its recognizability is variable. In this paper several definitions are given, such as dynamic feature space of patterns and dynamic recognizability of patterns. Necessity of multi-scale feature models of 3D target and rationality of using the general constraint for recognizing target image sequence are discussed. Based on these discussions, a multi-scale intelligent recursive recognizer (MUSIRR) is proposed for recognizing 3D moving targets, in which BP neural network and RBF neural network are the basic cells. During training, regular moment invariants of the multi-scale binary characteristic views of the target model are used as the pattern feature-vector. During recognition, the algorithm sufficiently uses reasonable restrictions of imaging process and target poses which are not changed acutely to achieve a good recognition ratio. Compared with the algorithms based on single-scale characteristic view models in references, the training of the MUSIRR algorithm is easy and needs less samples composed of the target characteristic view models. The algorithm can not only treat single frame images but also treat image sequence more effectively. The rationality and validity of the approach are proved by the results of massive simulation experiments on several kinds of aircrafts.